Abstract
In this study, a machine learning model for spatially predicting groundwater potential was developed using the XGBoost classifier, trained on 15 consistent geospatial predictors derived from topographic, hydrological, vegetation, and hydrogeological datasets. A total of 202 wells with measured groundwater discharge were used as training samples and converted into categorical groundwater productivity classes. The model integrates raster data from multiple sources, including elevation derived from SRTM, terrain indices (slope, TWI), hydrological distance measures, seasonal vegetation and moisture indices (NDVI, NDWI, BSI), and lithological and hydraulic properties derived from GLHYMPS. The interpretability of the model was assessed using SHAP analysis to quantify the contribution of each predictor to groundwater potential estimates. Furthermore, prediction uncertainty was quantified using the Shannon entropy of class probabilities, which enabled the identification of low-confidence zones associated with hydrogeological transitions. The results show that the model captures significant spatial patterns in groundwater potential, despite moderate prediction accuracy during spatial validation (F1 score ≈ 0.24). High uncertainty values are concentrated in structurally complex and lithologically heterogeneous zones. The proposed framework demonstrates a reproducible and transferable approach to integrating machine learning and geospatial analysis for mapping groundwater potential in data-deficient regions.